A fast method for finding charged-particle trajectories using unsupervised machine learning embedded in field-programmable gate arrays
Ashutosh V. Kotwal
We propose an algorithm, deployable on a highly-parallelized graph computing architecture, to perform rapid reconstruction of charged-particle trajectories in the high energy collisions at the LHC and future colliders. We use software emulation to show that the algorithm can achieve an efficiency in excess of 99.95% for reconstruction at high momentum with good accuracy. The algorithm can be implemented on silicon-based integrated circuits using FPGA technology. Since the algorithm requires no training, it represents a form of unsupervised machine learning. Our approach is potentially orders of magnitude faster than traditional computing, and may solve the challenge of the unaffordable computing cost of data processing at future colliders. It can also enable a fast trigger for massive charged particles that decay invisibly before reaching the muon detectors, as in some new-physics scenarios related to particulate dark matter. If production of dark matter or other new neutral particles is mediated by meta-stable charged particles and is not associated with other triggerable energy deposition in the detectors, our method would be especially useful for triggering on the charged mediators.